Marvell Wins $12.2B Google Chip Deal; Broadcom Drops 4.7%

The biggest AI chip story of the week belongs to Marvell, after the company won a custom silicon deal with Google and sent Broadcom into a sharp single-day slide.

Key points

• Marvell (MRVL) secured a custom AI chip deal with Google worth up to $12.2 billion in stock, directly challenging Broadcom’s standing as Google’s primary custom chip partner.
• Broadcom (AVGO) fell roughly 4.7% on the news, a notable move for a stock that recently projected $100 billion in AI revenue targets.
• Nvidia (NVDA) faces its own near-term catalyst: FQ2 earnings are approaching, with one analyst argument that the stock trades at a 50%-plus discount on a free cash flow basis, even after factoring in its $105 billion commitment to the OpenAI Ohio data center.
• Apple (AAPL) is holding up after a 10% correction but is facing margin pressure from a memory price upcycle that could weigh on iPhone profitability.
• LG Electronics is targeting 100,000 hours of robot training data by end of 2026, partnering with Nvidia to build its Robot Foundation Model ahead of a planned bipedal humanoid launch in Q1 2027.

What does the Marvell-Google deal mean for custom silicon?

The $12.2 billion figure is striking even by hyperscaler standards. Google has long relied on both its in-house TPUs and outside partners for custom inference and training accelerators, with Broadcom historically filling that outside role. Marvell’s win suggests Google is either diversifying its supply chain, found a price or performance edge with Marvell’s design, or both.

For Broadcom investors, the 4.7% drop reflects a real risk: custom AI chip revenue was a central pillar of the company’s bullish long-term revenue case. Losing Google, even partially, puts pressure on whether those projections hold. Broadcom has not yet publicly responded to the news, and the full terms of the Marvell deal, including timelines and performance milestones, are not yet clear from available reporting.

Marvell has spent several years repositioning itself around custom ASIC design for hyperscalers, so this deal is a validation of that strategy rather than a surprise pivot. The stock was already viewed as a credible Nvidia alternative for customers who want purpose-built inference silicon rather than general-purpose GPUs.

How does this fit into the Nvidia earnings setup?

Nvidia reports FQ2 results soon, and the framing heading in is unusually complex. The company carries an enormous headline commitment: up to $105 billion in credit and compute for an OpenAI data center in Ohio. That deal underscores Nvidia’s ambition to move beyond chip sales into infrastructure financing, but it also raises questions about capital deployment and balance sheet risk that investors will want answered on the call.

The FCF discount argument being made by at least one analyst is worth watching: if Nvidia’s near-term earnings are burdened by large upfront commitments while deferred revenue builds slowly, reported earnings could look optically weaker than the underlying business. How management frames that gap will likely set the tone for the stock’s next move.

Apple and memory: a quieter pressure point

Apple’s situation is less dramatic but worth tracking for anyone holding AAPL as an AI-adjacent position. The stock has recovered some ground after a 10% pullback, but a historic memory price upcycle is creating input cost pressure on iPhones. That matters because Apple’s hardware margin is the foundation its services multiple is built on. If DRAM and NAND prices stay elevated through the next iPhone cycle, the cost pressure will be real, even if Apple’s brand power limits how much it can be passed to consumers.

This connects to the broader memory story. The upcycle benefiting SK Hynix and Sandisk is a direct headwind for any company that buys large volumes of memory components. Apple buys an enormous amount.

LG and Nvidia: physical AI gets a timeline

LG’s announcement is a reminder that the AI investment cycle extends well beyond data centers. The company is using Nvidia infrastructure to generate both real and synthetic training data, targeting 100,000 hours by end of 2026, with a commercial humanoid robot planned for Q1 2027. Nvidia’s role as the compute backbone for physical AI training, not just cloud inference, adds another demand vector to its revenue story that does not show up cleanly in current GPU sales figures.

For investors, today’s session is a useful snapshot of how the custom silicon market is fracturing: Nvidia dominates general-purpose AI compute, but Marvell and Broadcom are fighting hard over the custom ASIC layer, and the outcome of that competition will shape hyperscaler capex allocation for years ahead.

This article is for informational purposes only and does not constitute investment advice.